summary or wc -l. A block is about 2,500 lines; the line count alone decides the split, never the number of files.

Inline. A small scope does not repay the cost of agents, which each rebuild the context and roughly double the review's token cost for the same findings. Do the analysis yourself: read references/agent-rules.md and the files (or the diff) with the one Bash call shown under the agent prompt below, and apply A1-A6 in a single pass under the same rules, confidence thresholds and output format. Do not re-report lint rows, and do not run inline above the limit because the code looks simple. Report Agents: 0 (inline).

Shards. Cut the files, in directory order, into n = ceil(lines / 2500) shards of about lines / n lines each; how many files a shard holds does not matter. Keep a directory in one shard where it fits, so that a class's header and source, and the models and views that reference each other, land together; split a larger directory by file. A file larger than a shard gets a shard to itself. Each shard agent applies all six sections to its shard only and may Grep the whole scope to trace a symbol. Run up to six shard agents at a time and the rest in a second wave. Cross-shard checks (ownership, model contracts, duplicated logic) are then best effort; say so in the report header.

Launch the agents in parallel, named after their shard (e.g. "Review shard 2/6: src/models"). Keep each prompt short; it contains only:

1. The absolute path of references/agent-rules.md and the section IDs to apply (A1-A6), plus the absolute path of references/qt-framework-checklist.md for A2 and A4 agents in framework mode

2. The root directory from the lint summary, the files in scope relative to it and, in diff scope, the git diff -U50 <range> -- <files> command for them

3. The lint summary rows for that agent's own files only (a shard agent never needs the other shards' rows), with the instruction "do not re-report these lines"

4. The instruction to load everything with one Bash call before analysing, and to return findings in the format at the top of the rules file. Each tool call re-sends the agent's whole context, so one batched read instead of one Read per file is the single largest saving on large scopes. Both forms print per-file line numbers for the file:line locations (cat -n on several files numbers them as one run, and git diff has no line numbers):

   # codebase scope
   cd <root> && cat <rules> && for f in <files...>; do echo "== $f"; cat -n "$f"; done
   # diff scope; numbers context and added lines with the new-file line
   cd <root> && cat <rules> && git diff -U50 <range> -- <files> | awk '/^@@/{split($3,a,/[+,]/);n=a[2];h=1;print;next} /^diff /{h=0} h&&/^[ +]/{printf "%5d %s\n",n++,$0;next} {print}'
   

Agents are read-only and must never edit files. In tools without agents, run each mission as whatever prompt chain or subprocess the tool supports; the contract is the same.

Phase 3: Consolidation

Merge lint and agent findings. Deduplicate on file+line+issue. When duplicates disagree on confidence, keep the block with the highest score and its trace; never average scores. Write the report in the format below.

Confidence scoring

Output format

Use exactly this structure. Lint findings are grouped by rule with one mitigation each; only deep-analysis findings get a full block.

## Qt Code Review Report

**Scope**: [diff: `git diff HEAD~1..HEAD` | files: <paths>]
**Files reviewed**: N
**Issues found**: N (M from lint, K from deep analysis)
**Agents**: N (shards ... | inline)
**Framework mode**: [on / off]

---

### Lint findings

| Rule | Count | Locations | Mitigation |
|---|---|---|---|
| LCY-2 QObject created with new but no parent | 3 | `main.cpp:12,40`, `widget.cpp:88` | Pass a parent or hold in a smart pointer. |
| ... | | | |

---

### Deep analysis findings

#### [D-NNN] <Short title>
- **File**: `path/to/file.cpp:42`
- **Category**: <A1-A6 mission name>
- **Confidence**: NN/100
- **Finding**: <what is wrong>
- **Trace**: <symbols followed, what was checked>
- **Mitigation**: <what to do, in prose — no code patches>

---

### Investigation targets (human verification needed)

#### [I-NNN] <Short title>
- **File**: `path/to/file.cpp:42`
- **Category**: <A1-A6 mission name>
- **Confidence**: NN/100
- **Finding**: <what is suspected>
- **Unverified because**: <what could not be confirmed>
- **How to verify**: <specific action for the reviewer>

---

### Summary

| Category | Lint | Deep | Investigate | Total |
|---|---|---|---|---|
| ... | N | N | N | N |
| **Total** | **M** | **K** | **I** | **N** |

Findings below confidence 60 are suppressed.

References

• references/agent-rules.md — agent missions A1-A6 and the agent output format (read by the agents, not the orchestrator)
• references/qt-review-checklist.md — universal rule IDs and rationale (human reference, not loaded during a review)
• references/qt-framework-checklist.md — framework rules (read by A2 and A4 agents in framework mode)
• references/qt-deprecated-classes.md — background for the DEP lint rules (human reference)
• references/lint-scripts/qt_review_lint.py — single-pass linter, --summary for grouped output, --diff=<range> for changed lines only, --framework for module rules